Agentic coding is the practice of using AI agents — not just AI assistants — to write, test, debug, and deploy software. The key difference: agents act autonomously within defined boundaries, while assistants wait for instructions.
| Level | Example | How It Works |
|---|---|---|
| Autocomplete | GitHub Copilot inline suggestions | Predicts the next few tokens |
| Chat Assistant | ChatGPT, Claude.ai | You describe, it generates code |
| Coding Agent | Claude Code, Cursor Agent Mode | Reads your project, makes multi-file edits, runs tests |
| Autonomous Agent | Devin, background agents | Plans and executes multi-step tasks independently |
Agent Loop — The cycle of: read context, decide action, execute, observe result, repeat. Agents loop until the task is done or they need human input.
Tool Use — Agents call tools (file read/write, terminal commands, web search, API calls) to interact with the real world, not just generate text.
Context Window — The amount of code and information an agent can "see" at once. Managing context is the primary skill of agentic coding.
System Prompt / CLAUDE.md — Instructions that configure the agent's behavior for your specific project. The better your context files, the better the agent performs.
MCP (Model Context Protocol) — An open standard for connecting AI agents to external tools and data sources. Think of it as USB for AI agents.
A2A (Agent-to-Agent) — Google's protocol for agents to communicate with each other. Still early but growing.
Search for a command to run...